物种敏感度分布的模型选择和最小样本量研究

Investigation of model choice and minimum sampling size for species sensitivity distribution

物种敏感度分布(SSD)是一种广泛用于生态风险评价和水质基准建立的统计分布模型。然而,模型的选择和样本量对SSD模型的构建和输出结果 HC5(即5%物种危害浓度)具有显著影响。收集了8种污染物对水生生物的急性毒性数据,每组数据均采用4种参数模型和2种bootstrap模型进行SSD的构建和HC5的估计,采用不同标准进行最优参数模型的选择。基于HC5值随样本量的变化情况,采用拐点分析法获得稳定SSD所需的最小样本量。参数模型选择的结果表明,3参数的Burr III模型总体上要优于其他3种2参数的模型(Log-normal,Log-logistic,Weibull),且部分污染物的最优模型与其选用的标准有关。对大部分污染物6种统计模型所推导的HC5相差2倍以上,表明模型选择是应用SSD模型过程中不确定性的重要来源。最小样本量的结果显示,非参数bootstrap和参数方法所需最小样本量的范围分别为5~12和5~13,中值均为8,即采用2种方法获得稳定SSD所需的最小样本量分别为12和13。数据量越大的毒性数据集,基于参数方法获得稳定HC5所需最小样本量也越大,因此建议在搜索获得污染物毒性数据后(筛选之前),应先推导出获得稳定HC5所需的最小样本量,以保证最终所构建的SSD在统计上是稳定的。

The paper is inclined to build up the species sensitivity distribution curve and estimate the volume size of hazardous concentrations at the 5 th centile( HC5) for each data set by collecting the 4 parameter models and 2 bootstrap models. And,then,to interpret their effects,we have managed to build up the data sets of the acute toxicities of the 8 substances to represent the range of classes with the different modes of action to the aquatic organisms. On the other hand,the said species sensitivity distribution( SSD) can be widely used to represent the ecological risk assessment statistical model and establish the water quality criteria. However,the model choice and sampling size tend to affect the development of the SSD model and the estimation of HC5 significantly. Therefore, the optimum parameter model has to be chosen to sum up the square deviations of the fitted model on the x-axis and y-axis by taking the Akaike information criteria( AIC) as a means for choice. Based on the variation of HC5 with the different sample sizes,the changing point analysis method has been adopted to gain the minimum sampling size needed for building up the stable SSD. The results of the parameter model chosen indicate that Burr III model is the optimized one for the toxicity data sets of ammonia,copper and PCBs whereas Weibull one is optimized for the South Louisiana crude oil. On the other hand,the log-normal model should be the best one for zinc,though the optimized model for the other 3 kinds of pollutants ought to be related to the chosen criteria. And,apart from the above mentioned,the HC5 derived from the 6 statistical models may differ more than trice for the most substances. All this may imply that that model choice should be taken as a significant origin of the uncertainty in assessing HC5. For example,the minimum sampling sizes for ammonia,cadmium,chlorpyrifos,cook bay crude oil,copper and zinc have to be 12,5,7,5,13 and 8 for the parameter method,respectively,whereas for the nonparametric bootstrap,the minimum sampling sizes should be 12,5,5,7,10 and 8,respectively and correspondingly.That is to say,the minimum sampling sizes demanded by the 2 above methods to obtain the stable SSD should be equal to 13 and 12,correspondingly. And,in turn,the greater the size of the parent sample,the higher the minimum sampling size has to be required based on the parametric method. Thus,the given research tends to imply that the minimum sampling size demanded to obtain the stable HC5 can be derived through searching for the contaminant toxicity data( before screening) to ensure that the final constructed SSD ought to be statistically stable and regular.